Reinforcement Learning Based Multi-robot Formation Control Under Separation Bearing Orientation Scheme

摘要

The multi-robot formation is a promising technology that endows robots with greater flexibility and cooperation capability. This paper aims at solving the separation-bearing-orientation scheme (SBOS) control problem of wheeled mobile robots (WMRs) based on the hybrid architecture of the improved deep deterministic policy gradient (DDPG) algorithm and the classical nonlinear formation control method. In particular, the trick of priority experience replay (PER) is utilized to improve the efficiency and stability of the whole training process. The proposed approach regards the controller gain hyperparameters of the three degrees of freedom in the WMR as the action space and develops the effective reward function to ensure the continuity of the velocity and acceleration while achieving the task. The simulation results of the two scenarios show that the proposed approach is available in performing various WMR-SBOS formation control tasks.